The Star in a Jar: AI in Nuclear Fusion and Plasma Physics

AI has produced important experimental results in plasma reconstruction, instability forecasting, magnetic shaping, and control. Those results do not establish a commercial fusion power plant. A controller that succeeds for a research discharge on one tokamak is evidence about a defined control problem; it is not proof of reliable electricity generation, qualified materials, a closed tritium fuel cycle, maintainable components, regulatory approval, or competitive lifecycle cost.

This distinction sharpens the engineering agenda. AI can help researchers run better experiments and close specific science and technology gaps, provided that learned components remain inside a measured, independently protected control architecture.

Define the experiment, not the headline

Start with a shot objective: maintain plasma current and position, follow a shape trajectory, estimate an unmeasured profile, avoid a tearing mode, predict a disruption, optimize heating, or compare operating scenarios. Specify the machine, plasma regime, pulse phase, diagnostics, actuator limits, control interval, termination rules, and scientific hypothesis.

Separate result classes:

• simulation performance is measured in a model;
• offline prediction uses recorded discharges without controlling hardware;
• hardware-in-the-loop testing includes real control interfaces and timing;
• experimental control acts on a research device during approved shots;
• reactor-relevant evidence addresses conditions and duration relevant to future plants;
• commercial generation requires an integrated facility that delivers usable electricity reliably and lawfully.

Moving between these classes requires new evidence. A laboratory milestone should not silently inherit the claim of the final class.

Build a discharge data lineage

Fusion data combine magnetic probes, flux loops, interferometry, spectroscopy, bolometry, cameras, temperature and density diagnostics, heating systems, gas injection, coil commands, protection events, equilibrium reconstructions, and operator annotations. Record calibration, coordinate system, sampling rate, clock alignment, valid range, preprocessing, missingness, saturation, and uncertainty for each channel.

Use immutable shot identifiers and preserve raw, conditioned, reconstructed, and model-ready layers. Link every training example to facility, campaign, machine configuration, wall condition, diagnostic version, plasma scenario, and code version. Prevent post-shot reconstruction or corrected labels from leaking into a historical real-time benchmark.

Cross-machine datasets need explicit semantic mapping. A signal with the same name may reflect a different diagnostic geometry, response, delay, or reconstruction method.

Choose the role of AI in the control stack

AI can serve several roles with different risks:

• state estimation infers quantities that are slow or unavailable in real time;
• forecasting estimates instability or disruption risk over a defined horizon;
• policy learning proposes actuator commands for an operator-specified objective;
• surrogate modeling approximates expensive plasma or engineering simulations;
• experiment planning ranks candidate settings under constraints;
• anomaly detection flags diagnostic, actuator, or subsystem behavior for review.

Do not combine all roles into an opaque “autonomous reactor.” Define the input-output contract, update rate, allowed operating domain, uncertainty behavior, and fallback for each component. Conventional estimation and control may remain the right choice for well-understood safety-critical loops.

Train against physics and operational limits

A learned controller is only as useful as its objective and environment. Include actuator voltage and current limits, slew rates, delays, sensor noise, power-supply dynamics, plasma termination conditions, wall-clearance limits, and relevant stability margins. Vary uncertain physical parameters across defensible ranges rather than adding arbitrary noise.

The experimentally demonstrated reinforcement-learning work on TCV used a simulator, sensor and actuator modeling, domain variation, and termination conditions before deployment. That is a control experiment, not a general claim that reinforcement learning can safely discover any reactor operating point.

Reward functions need review by plasma physicists, control engineers, machine operators, and protection specialists. Penalizing disruption is insufficient if a policy can maximize reward by entering an unqualified regime.

Verify from offline replay to approved shots

Use staged evidence:

1. unit-test preprocessing and signal semantics;
2. replay historical discharges with strict time causality;
3. compare with physics-based and conventional baselines;
4. stress calibration drift, dropout, latency, saturation, and off-normal states;
5. test in a high-fidelity simulator beyond the training distribution;
6. run on the target real-time hardware with measured worst-case latency;
7. conduct shadow operation during ordinary shots;
8. authorize bounded experimental control with independent abort criteria;
9. expand only after post-shot review.

Report negative and aborted experiments. Selecting only successful plasma pulses inflates reliability and hides the conditions that matter most.

Keep machine protection independent

Research plasma control and machine protection have different authority. A learned controller may pursue the approved shot objective inside a certified envelope. Independent interlocks and protection systems should enforce hard constraints, initiate mitigation, terminate a discharge, and protect magnets, vessel, plasma-facing components, heating equipment, cryogenics, and personnel.

The learned model should not redefine its safety envelope, suppress a protection signal, or approve its own deployment. Document command arbitration, watchdogs, stale-input behavior, compute failure, network isolation, manual takeover, safe actuator states, and post-event evidence.

Operators need an immediate way to return to a known controller or terminate the shot. For critical paths, keep deterministic timing and bounded resource use.

Evaluate prediction as an operational decision

For disruption or instability forecasting, define the event label, prediction horizon, minimum useful warning time, and the action triggered. Precision and recall alone are incomplete. Measure false alarms per pulse, missed events, time-to-event calibration, probability calibration, performance by scenario, and the cost of mitigation.

A predictor that warns after an intervention window is not operationally useful. One that alarms too often can waste shots, consume mitigation resources, and train operators to ignore it. Compare against locked baselines and evaluate on later campaigns, unseen configurations, and machine conditions.

For state estimates and controllers, report physical error, trajectory error, constraint violations, control effort, latency distribution, recovery, and robustness—not only aggregate reward.

Link plasma progress to the full plant problem

Future electricity requires much more than plasma control: materials under neutron exposure, heat exhaust, breeding and handling tritium, blankets, remote maintenance, availability, fuel cycle, turbine or conversion systems, licensing, waste management, supply chains, and economics.

ITER’s mission description states that it is an experimental device and will not generate electricity. Its planned fusion gain refers to power in the plasma, not net electricity delivered to the grid. DOE’s 2026 Fusion Science and Technology Roadmap describes milestones across multiple challenge areas and makes its timelines contingent on future partnerships and appropriations.

Keep these system boundaries visible when connecting fusion to the future of energy. A plasma-control paper should not be used as a power-market forecast.

Govern scientific models and claims

Assign ownership for diagnostics, training data, simulation, model code, real-time integration, shot authorization, machine protection, cybersecurity, and publication. Require reproducible experiment packages with configuration hashes, data lineage, approved operating domain, benchmark results, and model limitations.

Review dual-use and cybersecurity risks. Control software, facility layouts, operating data, and materials models may have access or export constraints. Segment research networks from protection systems, authenticate model artifacts, sign releases, restrict write paths, and audit changes.

Public claims should name the machine, method, experimental class, measured result, and limitation. Avoid “solved fusion,” “limitless energy,” and implied commercial dates unless an integrated evidence base supports them.

Measure the work that AI actually improves

Useful KPIs include:

• percentage of diagnostic channels with current calibration and lineage;
• reconstruction error and uncertainty by plasma regime;
• disruption forecast recall at a usable horizon and false alarms per shot;
• probability calibration and performance on later campaigns;
• shape, position, current, or profile tracking error;
• hard-constraint and protection-system interventions;
• real-time deadline misses and worst-case inference latency;
• successful recovery or fallback rate;
• shots saved, scientific information gained, and operator workload;
• reproducibility across repeated shots and independent teams;
• cross-machine transfer loss;
• model-caused aborted shots or equipment events.

Commercial metrics such as plant availability, net electricity, maintenance interval, tritium self-sufficiency, and levelized cost belong to integrated plant evidence. Do not substitute plasma reward or fusion energy in a pulse for them.

Anticipate failure modes

Design for the failures most likely to defeat a convincing demo:

• clocks drift and diagnostic channels are misaligned;
• a reconstruction used as “truth” changes after training;
• a model learns facility or campaign artifacts;
• simulator mismatch appears in an unmodeled regime;
• sensor saturation resembles a stable state;
• latency exceeds the control interval;
• a predictor is calibrated on too few disruptions;
• an optimizer exploits an incomplete reward;
• distribution shift follows wall conditioning or hardware change;
• a learned controller fights a conventional loop;
• operators cannot understand or override command arbitration;
• compute failure leaves an actuator command stale;
• publication selects only successful shots;
• a research milestone is marketed as imminent commercial power.

For each, define detection, an independent limit, safe response, responsible role, evidence capture, and a test before the next campaign.

Roll out by experimental authority

Begin offline on a narrow, measurable task with high-quality diagnostics and an established baseline. Freeze a temporal test set, publish failure slices, and demonstrate that the model adds value under latency and noise constraints.

Move to shadow mode on the target facility. Validate timing, integration, operator displays, cybersecurity, and divergence from the active controller. Let the machine owner and protection authority approve a bounded experimental envelope.

Then run supervised shots with conservative limits, immediate fallback, and post-shot review. Expand regimes one dimension at a time. Cross-facility transfer should restart validation rather than assume equivalence. Maintain a model-independent controller, exportable evidence, reproducible software environment, and rollback throughout.

The same energy discipline discussed in energy-aware computing applies here: computation has value when its measured scientific benefit justifies its hardware, operational, and verification burden.

Source notes

Source status was checked on 2026-07-30. The peer-reviewed TCV study, “Magnetic control of tokamak plasmas through deep reinforcement learning”, reports experimental control of specified plasma shapes and currents; it does not demonstrate a commercial reactor. DOE’s summary of reinforcement learning for tearing-instability avoidance describes research supported at DIII-D and KSTAR. ITER’s project milestone record reports that the ITER plasma control system was operated on KSTAR in March 2026 as a system test. ITER’s own mission description states that ITER will not convert fusion heat into electricity and distinguishes plasma gain from whole-plant electricity. DOE’s finalized 2026 Fusion Science and Technology Roadmap announcement is a national strategy whose activities and timelines remain contingent; it is not evidence that commercial generation has been achieved.